SnowPro Specialty: Gen AI Question 58
Single answerUnstructured data use casesA retail company stores thousands of product manuals, warranty PDFs, and installation guides in a Snowflake stage. The support team wants a GenAI assistant that can answer customer questions using the latest document content, while keeping the source files in Snowflake and minimizing data movement. Which approach is the MOST appropriate?
- A
Load each PDF into a structured table by manually converting every document into relational columns, then use only SQL filters to answer user questions.
- B
Use Snowflake support for unstructured data to access the staged files, extract document text for indexing and retrieval, and ground the model's responses on relevant passages from those documents.
- C
Export all staged files from Snowflake into an external vector database because Snowflake cannot participate in retrieval workflows for unstructured documents.
- D
Train a custom large language model from scratch on the manuals and guides so the model memorizes the document contents and no retrieval step is needed.
Show answer and explanation
Correct answer: B
Explanation
The key requirement is to answer questions over changing document content while keeping files in Snowflake and minimizing unnecessary movement. That points to a retrieval-based architecture over unstructured data, not manual relational remodeling, full export to another platform, or training a model to memorize documents. In Snowflake, unstructured data can be stored in stages and used in AI workflows where text is extracted from files and then used for search, retrieval, and grounding. This reflects common GenAI best practice: use RAG for enterprise document question answering so responses are based on current source content rather than static model memory. Candidates should recognize that unstructured data use cases in Snowflake often involve staged files, document processing, and retrieval-centric patterns rather than forcing all content into structured schemas or relying on model training alone.
- A. Incorrect.
This is not the most appropriate approach. While some document metadata or extracted fields can be modeled in structured tables, manually converting every PDF into relational columns is operationally heavy, brittle, and poorly suited to long-form document question answering. It also ignores the core value of unstructured data workflows, where the original files remain in object storage or stages and are processed for search and retrieval.
- B. Correct.
This is correct. For a GenAI assistant over manuals and guides, the practical pattern is retrieval-augmented generation (RAG): keep the files in Snowflake-managed or external stages, process the unstructured content to extract text, create searchable/indexable representations, retrieve relevant passages, and use those passages to ground responses. This supports fresher answers than model-only approaches and aligns with Snowflake best practices for working with unstructured data and document-based AI use cases.
- C. Incorrect.
This is incorrect because it assumes Snowflake cannot support retrieval workflows involving unstructured documents. Snowflake supports unstructured data use cases and can participate in document processing and retrieval patterns without requiring all data to be moved out to another platform. External tools may be used in some architectures, but exporting everything is not inherently required and conflicts with the scenario's goal of minimizing data movement.
- D. Incorrect.
This is incorrect because training a custom LLM from scratch is expensive, slow, and unnecessary for most enterprise document QA scenarios. It also creates a freshness problem: new manuals or revised warranty terms would require retraining or fine-tuning cycles. In practice, grounding a model with retrieved document context is the preferred approach for enterprise knowledge assistants over changing unstructured content.